SSHF-DTI: Leveraging structural similarity and hierarchical features through a fusion network for drug-target interaction prediction.

Predicting drug-target interactions (DTI) and binding affinities (DTA) is essential for drug discovery, but experimental methods remain costly and time-consuming. While deep learning approaches have improved prediction performance, many existing models rely on single-source data and lack integration of cross-domain features, limiting their generalization. This study aims to develop a more robust and generalizable model for DTI and DTA prediction. We propose SSHF-DTI, a model that integrates structurally similar information and multi-source substructure features to effectively preserve chemically meaningful molecular fragments and capture hierarchical feature dependencies. Structural similarity is incorporated through data enrichment based on Tanimoto coefficient evaluation of Morgan fingerprint similarity. This hybrid architecture combines transformer and convolutional components and further achieves hierarchical feature fusion, thereby significantly optimizing model performance improvement. Compared with baseline methods, SSHF-DTI achieved improvements of 0.031 and 0.147 in ROC-AUC and PR-AUC respectively on the Davis dataset. It also demonstrated strong generalization in drug-drug interaction (DDI) tasks and showed high sensitivity in distinguishing fine-grained molecular structural features affecting binding affinity. SSHF-DTI provides a powerful, generalizable framework for DTI/DTA/DDI prediction, capable of capturing complex hierarchical feature interactions. It shows promise for supporting drug discovery and virtual screening applications.